Follow the leader: Herding behavior in heterogeneous populations

PHYSICAL REVIEW E 91, 052804 (2015)
Follow the leader: Herding behavior in heterogeneous populations
Guillem Mosquera-Do˜nate and Mari´an Bogu˜na´
Departament de F´ısica Fonamental, Universitat de Barcelona, Mart´ı i Franqu`es 1, 08028 Barcelona, Spain
(Received 23 December 2014; revised manuscript received 31 January 2015; published 11 May 2015)
Here we study the emergence of spontaneous collective leadership in large populations. In standard models of
opinion dynamics, herding behavior is only obeyed at the local scale due to the interaction of single agents with
their neighbors; while at the global scale, such models are governed by purely diffusive processes. Surprisingly,
in this paper we show that the combination of a strong separation of time scales within the population and a
hierarchical organization of the influences of some agents on the others induces a phase transition between a
purely diffusive phase, as in the standard case, and a herding phase where a fraction of the agents self-organize
and lead the global opinion of the whole population.
DOI: 10.1103/PhysRevE.91.052804
PACS number(s): 89.75.Fb, 87.23.−n, 05.40.Ca, 89.75.Hc
Humans are unpredictable but collective human behavior
can be predicted. Although apparently contradictory, this
statement is the main hypothesis of sociophysics [1–3]. By
analogy with nonequilibrium statistical mechanics, the hope
is that collective social phenomena can be treated as emerging
properties of systems of interacting agents that depend on a
few fundamental features of the microscopic interaction laws,
rather than on the idiosyncratic character of single individuals.
This hope has encouraged the study of social dynamics using
the tools and models that statistical physics has been developing for the last 50 years or so [4]. Opinion dynamics is one
of the better examples of the application of this approach. The
aim here is to understand the opinion of a population of agents
and the rules that govern transitions between different opinion
states as a response to social influence, or the tendency of
people to become like those they have social contact with [5].
Models of opinion dynamics typically show consensus
states, where the dynamics is frozen. In many cases, as in
the voter [6,7] or Sznajd models, the (weighted) ensemble
average opinion of the population is a conserved quantity. In
such cases, the dynamics of the stochastic average opinion
is governed by a purely (nonhomogeneous) diffusive process
without any drift, which eventually leads the system to one of
the possible consensus states. It is therefore difficult to imagine
how collective leadership can emerge in this context. In this
paper we show that leadership can, in fact, arise spontaneously
in a subset of the population when there is strong heterogeneity
in the time scales of the agents coupled with a hierarchical
organization of their influence. Heterogeneity of time scales
is present, for instance, in speculative markets, where noise
traders who operate at the scale of minutes or hours coexist
with fundamentalists, who operate at the scale of weeks or
months. Interestingly, we reveal a pitchfork bifurcation that
separates a purely diffusive phase from a phase where the most
active agents lead the global state of the entire population. Our
results could shed light on the dynamics of financial crises and
other extreme events caused by humans.
The voter model was first introduced in 1973 to model
competition between species [6,7]. Ever since, it has been
one of the most paradigmatic and popular models of opinion
dynamics. Its simplicity, analytical tractability, and versatility
when it comes to introducing new mechanisms make it the
perfect model for studying many different phenomena in
both the natural and social sciences, from catalytic reaction
1539-3755/2015/91(5)/052804(5)
models [8,9] to the evolution of bilingualism [10] or US
presidential elections [11]. In its simplest version, the voter
model is defined as follows. There is a set of N interacting
agents, each endowed with a binary state of opinion (sell or
buy, Democrat or Republican, Windows or Mac, etc). For each
time step of the simulation, an agent is randomly chosen to
interact with one of their social contacts, after which that agent
copies the opinion of their contact.
Heterogeneity can be introduced within the population
through the activity rate of agents [12,13]. We assume that
agents are given intrinsic activity rates {λi }, which control the
frequency at which the agents interact with their social contacts
and, possibly, change their opinion. In numerical simulations,
this is equivalent to choosing the next active agent, say agent i,
with probability proportional to λi . The influence of one agent
on others can be modeled by the probability Prob(j |i) that
agent i copies the opinion of agent j when i is activated at rate
λi . When contacts take place according to a fixed social contact
graph with an adjacency matrix aij , this probability is given by
Prob(j |i) = aij /ki , where ki is the degree of agent i [14–17].
If a fully connected graph pertains (equivalent to a mean-field
description), this probability is simply Prob(j |i) = 1/(N − 1)
for j = i and zero otherwise.
The dynamics of the state of the system can be described
using a set of N dichotomous stochastic processes {ni (t)}
that take the value 0 or 1 depending on the opinion state
of each agent at time t. If we assume that all temporal
processes follow Poisson statistics, the stochastic evolution
of ni (t) after an increment of time dt satisfies the stochastic
equation [18,19]
ni (t + dt) = ni (t)[1 − ξi (t)] + ηi (t)ξi (t),
(1)
where ξi (t) is a random dichotomous variable that takes values
1 with probability λi dt,
ξi (t) =
(2)
0 with probability 1 − λi dt.
Notice that ξi (t) controls whether node i is activated during
the time interval (t,t + dt). If it is, the opinion of a neighbor
will be changed according to Prob(j |i), so that
1 with probability N
j =1 Prob(j |i)nj (t),
ηi (t) =
0 with probability 1 − N
j =1 Prob(j |i)nj (t).
(3)
052804-1
©2015 American Physical Society
˜
´ BOGUN
˜A
´
GUILLEM MOSQUERA-DONATE
AND MARIAN
PHYSICAL REVIEW E 91, 052804 (2015)
This equation implies the existence of a global conserved
magnitude [20,21] related to the eigenvector φ(i) of eigenvalue
1 of Prob(j |i); that is, the solution of the equation
φ(i)Prob(j
|i) = φ(j ). Indeed, by multiplying Eq. (4) by
i
φ(i)/λi and summing over all agents, the right-hand side of the
equation vanishes. Therefore, the weighted ensemble average
of the population:
≡
N
φ(i)
i=1
λi
ρi (t) =
N
φ(i)
i=1
λi
ρi (0)
(5)
is conserved by the dynamics and thus it is a function only of
the initial conditions; as we show above in the right-hand hand
side of Eq. (5). This fact can be used to evaluate the probability
of the final fate of a realization of the dynamics. For instance,
the probability of ending
up absorbed in the “1” consensus
state is given by just / i φ(i)/λi . Note that in the case of a
fixed social contact graph, φ(i) = ki .
The results presented so far are valid for an arbitrary distribution of individual rates λi . However, while the existence
of the conserved quantity tells us about the final fate of the
system, the approach of the system towards such fate can be
very different from the standard voter model when there is a
strong separation of time scales coexisting in the system (as
we say above, as in speculative markets with noise traders and
fundamentalists). To shed light on this problem, hereafter we
analyze a simple model with a population segregated into two
groups: a fast group of size Nf , operating at rate λf ; and a slow
one of size Ns , operating at rate λs , with λf > λs . Aside from
heterogeneity in their time scales, agents in a real population
are also heterogeneous in terms of their influence on others.
To model this effect, we assume that the probability of agent
i copying the opinion of agent j is a function of the rate at
which agent j operates, that is
f (λj )
Prob(j |i) = N
,
i=1 f (λi )
(6)
where f (λ) is an arbitrary function that represents the reputation of agents operating at rate λ as seen by the population. Note
that in this case, φ(i) = f (λi ). When f (λ) is a monotonically
increasing function, the influence of agents is hierarchically
organized, with fast agents having higher reputations and thus
being copied more frequently, by both fast and slow agents.
Similar to [22], in this work we use f (λ) = λσ .
1
fraction of agents in state "1"
In principle, ηi (t) should be realized only when ξi (t) = 1.
However, due to the particular form of Eq. (1), the value of
ηi (t) is only relevant when ξi (t) = 1. Therefore, we can safely
consider ξi (t) and ηi (t) as statistically independent random
variables.
Equation (1), supplemented with the definitions of variables
ξi (t) and ηi (t), represents the exact stochastic evolution of the
system. For instance, the ensemble average of the opinion of
agent i, ρi (t) ≡ ni (t), can be evaluated by taking the average
of Eq. (1) first over the variables ξi (t) and ηi (t), and then
over the ensemble. This program leads to the exact differential
equation
⎡
⎤
N
dρi
= λi ⎣
Prob(j |i)ρj − ρi ⎦ .
(4)
dt
j =1
f
0.5
0
0
1
5
10
3
=3x10
f
0
0
1
5
0.5
=10
f
10
15
10
15
s
20
15
0.5
0
0
=10
25
3
s
20
25
20
25
4
s
5
time (in units of
-1
)
s
FIG. 1. Single realizations of the stochastic evolution of the
fraction of agents in the 1 state for a two-compounded heterogeneous
system with Nf = 1000 fast and Ns = 4000 slow agents and different
time scales, λf = 103 λs (top), λf = 3 × 103 λs (middle), and λf =
104 λs (bottom). Simulations are generated as follows: at each time
step, with probability p = (1 + λs Ns /λf Nf )−1 a fast agent i (chosen
randomly within the population of fast agents) is activated. Then, with
probability given in Eq. (6) she copies the opinion of agent j . With
complementary probability 1 − p a slow agent is activated and copies
the opinion of another agent with Eq. (6). Time is then increased as
t → t + (λf Nf + λs Ns )−1 .
Figure 1 shows particular realizations of the process in
a system consisting of a small group of fast agents, Nf =
1000, and a large group of slow agents, Ns = 4000. In this
particular example, we set σ = 1 and different time scales
λf /λs . When the separation of the time scales between the
two groups is not very important, the global dynamics is purely
diffusive; as in the standard voter model (top panel of Fig. 1).
However, when the separation of time scales exceeds a certain
critical value, the behavior changes completely. Periods of
quasiregular increase and decrease alternate, and are suddenly
broken by sharp peaks. Although the system ends up in one
of the two absorbing states, the peculiar pathway followed to
reach consensus is not observed in the standard voter model.
To understand this phenomenon, in Fig. 2 we show the
temporal evolution of the two groups separately. From this
figure, it is clear that the anomalous behavior we observe in
Fig. 1 is the result of the highly differentiated dynamics of the
fast and slow agents. Due to the huge differences between the
time scales, from the perspective of the fast group, the slow
agents seem to be frozen in their state. However, due to the
monotonic increasing form of function f (λ), the effect of slow
agents on the dynamics of fast ones is small. In this situation,
fast agents evolve as in the simple voter model until they reach
one of their consensus states. Nonetheless, in contrast to what
we observe in the simple voter model, this consensus state
is not an absorbing one. Indeed, despite the low probability
of a fast agent copying a slow one, the time scale of the fast
agents is short enough for just this interaction to occur many
times during the evolution of the system. When such an event
052804-2
FOLLOW THE LEADER: HERDING BEHAVIOR IN . . .
PHYSICAL REVIEW E 91, 052804 (2015)
fraction of agents in state "1"
1
0.5
fast agents
0
5
1
10
15
random variables; so the central limit theorem can be invoked.
As a result, we conclude that the stochastic evolution of the
≡ [f (t),s (t)] can be described by a Langevin
vector (t)
equation. In particular, for the fast group dynamics, we can
write
df (t)
= Af [(t)]
+ Df [(t)]ξ
(8)
f (t),
dt
where ξf (t) is Gaussian white noise. The drift and diffusion
terms are defined respectively in terms of the infinitesimal
moments as
0.5
Af =
slow agents
0
10
15
time (in units of
)
s
Af = αf s (s − f ),
occurs, a fast agent may adopt an opposite opinion from a slow
outsider, thus introducing some noise into the small subsystem
and preventing it from becoming trapped in the consensus
state. In other words, the absorbing boundary is replaced by
a reflecting one. The same noise induced by slow agents can
make the group of fast agents change abruptly to the opposite
state, thereby providing the system with an effective two-state
dynamics, as can clearly be seen in the top panel of Fig. 2.
At the same time, from the perspective of the slow group,
fast agents spend long periods of time in the consensus states.
Again, due to the monotonic increasing form of function f (λ),
slow agents have a greater tendency to copy the opinion of
fast agents who, being quasifrozen in the consensus state, act
as a constant drift that pulls the opinion of the slow agents
towards the opinion of the fast ones. We can interpret this
as the group of slow agents becoming a herdlike group that
follows the leadership of the group of fast agents. However, this
behavior is not observed across the whole range of parameters
and, at this point, it is unclear whether it appears suddenly at a
critical value or is a crossover effect interpolating continuously
from the diffusive behavior of the standard voter model to the
herding behavior we observe in Fig. 1.
The existence of the conserved quantity implies that the
dynamics cannot be completely understood only in terms of
Eq. (4), as that equation does not contain any information
regarding the noise in the system. So we are forced to develop
a theory that includes second-order terms in the dynamics. To
do this, we take advantage of the homogeneity within each
group of agents and define the instantaneous average opinion
state of each group as
1 ni (t);
Nf i∈fast
s (t) ≡
1 ni (t).
Ns i∈slow
(7)
In the limit of large systems, f (t) and s (t) can be considered
as quasicontinuous stochastic processes in the range [0,1].
Furthermore, they result from the sum of a large number of
(9)
where f (t) ≡ f (t + dt) − f (t) [23]. These two terms
can be computed exactly using Eq. (1) and read
FIG. 2. (Color online) Evolution of the fraction of fast (top) and
slow (bottom) agents in state 1 of the same system as in Fig. 1. The
plots correspond to the supercritical phase with λf = 104 λs .
f (t) ≡
f (t)|(t)
[
f (t)]2 |(t)
, Df =
,
dt
dt
Df =
αf s
(s + f [1 + 2βf s − 2s − 2βf s f ]),
Nf
(10)
(11)
where we have defined
αf s =
λf
Nf f (λf )
and βf s =
.
1 + βf s
Ns f (λs )
(12)
Similar equations can be derived for the slow group by
replacing the index f ↔ s in the preceding equations.
When the separation of time scales is large, the state of
the slow group is perceived by the fast group as constant. In
this case, we can consider s in the previous equations as a
constant parameter. As a consequence, the diffusion term in
Eq. (11) does not vanish when f = 0, or f = 1 and the
system reacts at these points as it does in the presence of
a reflecting barrier. Therefore, the system has a well-defined
steady state controlled by an effective potential that, up to a
constant value, takes the form [23]
Af
Veff (f ) = ln Df − 2
df .
(13)
Df
This potential has a single extremum at approximately f∗ =
s , which changes from being a minimum to a maximum when
2
f (λf )
> Ns .
f (λs )
(14)
When this condition is met, the combination of a maximum
at f ≈ s with the two reflecting barriers at f = 0 and
f = 1 transforms the effective potential into a double-well
potential with a barrier at f ≈ s . This defines a pitchfork
bifurcation that separates a diffusive phase, in which the fast
group is dragged down by the slow one, from a herding phase,
in which the fast group effectively behaves as a two-state
system; jumping from one state to the other as in an activated
process. This is illustrated in Fig. 3, where we show the
effective potential when s = 0.5 in the two cases, along with
examples of realizations of the slow and fast group dynamics.
It is important to stress that condition Eq. (14) is independent of
the value of s . Therefore, even though the effective potential
052804-3
˜
´ BOGUN
˜A
´
GUILLEM MOSQUERA-DONATE
AND MARIAN
0
6
3
s
Veff( f)
Veff( f)
=10
f
4
2
-2
-4
=10
f
0
0.5
1
f
1
0.5
0
0
-6
fraction of users in state "1"
fraction of users in state "1"
0
PHYSICAL REVIEW E 91, 052804 (2015)
0
0.5
4
s
1
f
1
0.5
5
time
10
0
5
10
15
time
FIG. 3. (Color online) The upper plots show the effective potential below and above the critical point in the system represented in
Fig. 1 for a fixed value of s = 0.5. The lower plots show typical
realizations of the evolution of fast (red) and slow (blue) agents in
both cases.
in a quite straightforward way, even for very large populations,
in particular when the exponent σ > 1.
In this paper we present the minimal mechanisms that give
rise to the emergence of leadership and herding behavior in a
population of interacting agents: a strong separation of time
scales coupled with some form or hierarchical organization
of the influence of some agents over the others. Despite
the simplicity of the toy model that we use in this work,
the mechanisms are general enough to be extrapolated to
more complex and realistic situations. For instance, the
simple segregation of the population into only two groups is
probably an assumption too strong, given the heterogeneity of
human populations. Nonetheless, this assumption is not really
necessary; although mathematically more involved, it can be
shown that the same phenomenology occurs in systems with a
strongly heterogeneous distribution of activity rates, as usually
found in financial markets [24]. The hierarchical organization
can also be induced by different mechanisms, such as a
hierarchical organization of the network of contact among the
agents, formed of a core of well-interconnected agents and a
periphery of agents that are mainly connected to the core, as
can be observed in many real complex networks [25]. Finally,
one could also argue that the influence that a group of agents
has on the others is itself a stochastic process. In our case,
such a scenario could easily be modeled by assigning some
stochastic dynamics to the parameter σ . This is particularly
interesting since, as the transition is effectively discontinuous,
the dynamics would be a mixture of purely diffusive periods,
during which σ is such that the condition in Eq. (14) is
violated, and periods with strong herding behavior the rest of
the time.
is modified as s slowly evolves, its qualitative shape (whether
it shows a minimum or a maximum near s ) is not modified.
We should also note that, while this transition is not a
true phase transition (as it disappears in the thermodynamic
limit Ns 1), for finite systems it behaves effectively as a
first-order phase transition. This is due to the abrupt change,
at the critical point Eq. (14), of the effective potential Eq. (13),
which changes from having a single stable minimum to being
a bistable potential, exactly as in first order phase transitions.
Moreover, the strong separation of time scales we find in
some real systems, such as speculative markets (which can
be of order λf ∼ 104–5 λs ), coupled with a growing preference
function f (λ) ∼ λσ , can result in condition Eq. (14) holding
This work was supported by a James S. McDonnell
Foundation Scholar Award in Complex Systems; the ICREA
Academia prize, funded by the Generalitat de Catalunya; the
MINECO Project No. FIS2013-47282-C2-1-P; the AGAUR
Grant No. 2014SGR608; and by the European Commission
within the Marie Curie ITN “iSocial”, Grant No. PITN-GA2012-316808.
[1] P. Arnopoulos, Sociophysics: Cosmos and Chaos in Nature and
Culture (Nova Science, Hauppauge, NY, 1993).
[2] S. Galam, Sociophysics: A Physicist’s Modeling of Psychopolitical Phenomena (Springer, New York, 2012).
[3] P. Sen and B. K. Chakrabarti, Sociophysics: An Introduction
(Oxford University Press, Oxford, 2013).
[4] C. Castellano, S. Fortunato, and V. Loreto, Rev. Mod. Phys. 81,
591 (2009).
[5] L. Festinger, Schachter, and K. Stanley; Back, Social Pressures
in Informal Groups: A Study of Human Factors in Housing
(Stanford University Press, Stanford, CA, 1950).
[6] P. Clifford and A. Sudbury, Biometrika 60, 581 (1973).
[7] R. A. Holley and T. M. Liggett, Ann. Prob. 3, 643 (1975).
[8] J. W. Evans, Langmuir 7, 2514 (1991).
[9] J. W. Evans and T. R. Ray, Phys. Rev. E 47, 1018 (1993).
[10] X. Castell´o, V. M. Egu´ıluz, and M. San Miguel, New J. Phys. 8,
308 (2006).
[11] J. Fern´andez-Gracia, K. Suchecki, J. J. Ramasco, M. San Miguel,
and V. M. Egu´ıluz, Phys. Rev. Lett. 112, 158701 (2014).
[12] N. Masuda, N. Gibert, and S. Redner, Phys. Rev. E 82, 010103
(2010).
[13] J. Fern´andez-Gracia, V. M. Egu´ıluz, and M. San Miguel, Phys.
Rev. E 84, 015103 (2011).
[14] V. Sood and S. Redner, Phys. Rev. Lett. 94, 178701 (2005).
[15] K. Suchecki, V. M. Egu´ıluz, and M. S. Miguel, Phys. Rev. E 72,
036132 (2005).
[16] F. Vazquez and V. M. Eguiluz, New J. Phys. 10, 063011
(2008).
[17] V. Sood, T. Antal, and S. Redner, Phys. Rev. E 77, 041121
(2008).
[18] M. Catanzaro, M. Bogu˜na´ , and R. Pastor-Satorras, Phys. Rev. E
71, 056104 (2005).
[19] M. Bogu˜na´ , C. Castellano, and R. Pastor-Satorras, Phys. Rev. E
79, 036110 (2009).
052804-4
FOLLOW THE LEADER: HERDING BEHAVIOR IN . . .
PHYSICAL REVIEW E 91, 052804 (2015)
[20] K. Suchecki, V. M. Egu´ıluz, and M. S. Miguel, Europhys. Lett.
69, 228 (2005).
´ Serrano, K. Klemm, F. Vazquez, V. M. Egu´ıluz, and
[21] M. A.
M. S. Miguel, J. Stat. Mech. (2009) P10024.
[22] A. Chakraborty and S. S. Manna, Phys. Rev. E 81, 016111
(2010).
[23] G. W. Gardiner, Handbook of Stochastic Methods for Physics,
Chemistry and the Natural Sciences (Springer, Berlin, 2004).
[24] X. Gabaix, P. Gopikrishnan, V. Plerou, and H. E. Stanley, MIT
Working Paper Ser. 1, 3 (2003).
[25] P. Csermely, A. London, L.-Y. Wu, and B. Uzzi, J. Complex
Networks 1, 93 (2013).
052804-5